Transcript .ppt
ACT-R 5.0 and ACT-R/PM Michael D. Byrne Department of Psychology Rice University Houston, TX 77005 [email protected] http://chil.rice.edu Overview RPM and 5.0 RPM opportunities and future work 2 Buffer syntax Cognition parallels Activation sourcing Compatibility issues EMMA Top-down vs. bottom-up attentional control Visual object synthesis Lotsa other stuff ACT-R/PM (under 4.0) 3 5.0 Buffer Syntax LHS now consists entirely of testing the state of various “buffers” Goodbye to “time now”! Elimination of “!send-command!” syntax 4 Goal buffer Retrieval buffer PM state buffers (e.g., motor-state) Visual-location and visual (object) buffers Aural-locaiton and aural (object) buffers Use “+” syntax on RHS to send commands ACT-R 4.0 vs. 5.0 (p look-label =goal> isa do-menu target nil =loc> isa visual-location time now screen-x LOWEST attended nil =vis> isa module-state module :vision modality free =mot> isa module-state module :motor modality free ==> !send-command! :VISION move-attention :location =loc :scale WORD !send-command! :MOTOR move-cursor :loc =loc ) 5 (p look-label-5 =goal> isa do-menu target nil =visual-location> isa visual-location =visual-state> isa module-state modality free =motor-state> isa module-state modality free ==> +visual> isa visual-object screen-pos =visual-location +manual> isa move-cursor loc =visual-location ) Ramifications Cleaner syntax (yay!) More consistent No way to confuse RPM calls and retreivals Issues Restricts motor flexibility No parallel to “time now” LHS test on visual-location 6 Each command is a chunk type, therefore fixed # of arguments The PREPARE command takes a variable number of arguments Under 5.0, can only request an action on a buffer in the RHS LHS is only for tests of a buffer Two Productions (p look-label-5 =goal> isa do-menu target nil =visual-location> isa visual-location =visual-state> isa module-state modality free =motor-state> isa module-state modality free ==> +visual> isa visual-object screen-pos =visual-location +manual> isa move-cursor loc =visual-location ) 7 (p find-label-5 =goal> isa do-menu target nil ==> +visual-location> isa visual-location screen-x lowest attended nil ) Visual-location Testing Thus, the “find-and-shift” idiom has to be split across two productions This affects timing—old shift time was 185 ms (one 50 ms production, one 135 ms shift) This affects state control 8 An extra production required at each step Attention shift latency dropped to 50 ms (why not 85?) Both of those productions will match, so now we need to be more restrictive with conditions The (current) solution: “buffer stuffing” Visual-locations automatically “stuffed” into the =visual-location> buffer Default is newest & furthest left (lowest screen-x) Buffer-stuffing Issues This creates one other problem: Is the improvement in syntax worth breaking the idiom? Discussion: We could make the =visual-location> and =aural-location> buffers “instant” buffers 9 Display updates cause an implicit attention shift to the currently-attended location (the “blink” problem) Not consistent with buffer stuffing That is, not requiring RHS call-out Breaks parallel syntax (bad) Fixes timing issue and blink issue (good) Improves code-level compatibility with 4.0 (good) Would models be easier or harder to understand? Cognition-PM Parallels 5.0 makes the declarative memory system and the visual/audio systems look very much alike But for PM requests, it is possible for a production to check whether a request is in progress 10 Set up a request for information on the RHS Get it back in a buffer Asynchronous For example, by testing the =visual-state> So, should there be a =retrieval-state> ? Note that it is possible to set up a retrieval and harvest it in two productions, but vision/audio requires three Activation Sources Under 4.0, the slots of the currently attended visual object (and the currently attended sound) were activation sources This enabled ACT-R to rapidly answer questions like “what color is the thing you’re looking at?” Thus, it is retrieved very quickly Should properties of attended object be highly accessible? This has been removed for 5.0 11 color slot of object was activation source ? Backward Compatibility Issues How many RPM models based on 4.0 will break under 5.0? Implementation 12 In principle, very few: “time now” could just be translated to a buffer test However, find-and-shift idiom will have some trouble being translated 5.0 makes a lot of under-the-hood changes that render it not backward-compatible at the code level Maintaining one version of RPM is taxing enough, I don’t know about maintaining two Should all future development of RPM assume ACT-R 5.0? EMMA I keep saying that if people have ideas about extending RPM, by all means bounce it off me and we’ll see how it goes Dario Salvucci’s Eye Movements and Model of Attention (EMMA) extension to the Vision Module 13 This has finally happened! Separates attention shifts and eye movements Now part of the RPM 2.0 release (should work with 5.0 but I’m not sure yet) Dario wrote the original, and Dario and I hammered out a new version Still some unresolved issues Bottom-up vs. Top-down Control of Attention Attentional control in RPM (under 4.0) is entirely topdown Buffer stuffing gives some modicum of bottom-up attentional control How should this work? 14 Current literature on top-down vs. bottom-up control is mixed Best guess seems to be that top-down settings override bottom-up when present, but there isn’t always top-down Something like the Wolfe model might work, except that isn’t fleshed-out enough to implement I have a grad student working on this What is the identity of the green LETTER? 5 7 2 G 8 T 3 4 3 15 9 3 6 1 Visual Object Synthesis 16 Scales are all defined for text (phrase, word, letter) but not for other objects How should it work more generally? Use angle-based scale? Other Issues We Could Discuss Number of finsts Targeting nothing still isn’t really worked out Visual guidance constraints on aimed movements are not really enforced 17 Should they be? If so, how? Movement noise Spatial cognition